Safety Data Was Never a Cost. AI Just Proved It.

How pharmacovigilance, the industry's classic cost center, becomes its revenue engine

Safety Data Was Always More Than Paperwork

For most of the working life of pharmacovigilance scientists, the safety of medicines has been treated as paperwork. Side effect reports arrive, are read, coded, filed and sent to regulators on time. Success looks like nothing happening. A department whose best year is identical to its worst is one nobody invests in.

That was always wrong. Safety data is the only record that keeps growing about what a medicine does in ordinary people, including the pregnant, the elderly and the very ill, who are kept out of trials. It is the largest evidence base in the industry, and we have been storing it rather than reading it.

From Lucky Accidents to a Repeatable Method

Look at what came out of it by accident. Minoxidil was a blood pressure drug until patients grew hair. Sildenafil disappointed as a heart drug and astonished elsewhere. Each discovery hung on one alert doctor noticing something odd. Luck does not scale.

It no longer has to. The method is simple. Take the records of people who have already taken the drug, and write down the rules for a fair comparison before looking at the answer: who counts, who they are compared against, and what would prove it wrong. 

Done honestly, it behaves like a trial nobody had to run. In the records of over a million patients, people taking semaglutide were diagnosed with Alzheimer's less often, and people on similar drugs were hospitalized for alcohol problems less often. Both findings came out of data collected for safety.

Redesigning Products and Earning Approvals

The same data redesigns products. The American limit of 325 mg of paracetamol in prescription painkillers exists because liver injury reports piled up until the pill had to change. OxyContin was reformulated to resist crushing because surveillance showed how it was misused. Side effect reports became engineering instructions.

It can earn approvals. In 2019, palbociclib was approved for men with breast cancer in ordinary hospital settings, because there would never be enough such men for a trial. In 2021, tacrolimus was approved for lung transplant patients in a registry, with the FDA stating that a well-designed study of real patients against a fair comparator can count as proper evidence.

The Evidence Already Exists. We Just Aren't Using It.

Here is what that looks like next. A woman of thirty-two asks whether her medicine is safe in pregnancy and is told the data are limited, when thousands of women have already taken it in pregnancy, and their outcomes sit in records nobody has assembled. A man of eighty-two takes eight medicines in a combination no trial ever tested, and the interaction that will hospitalize him next spring is already visible in people like him. A child has a disease affecting four hundred people worldwide, so no trial can be large enough, and their experience is the only evidence there will be.

Regulators Move First — And Stumble

None of it was practical until machines learned to read the messy sentence a nurse typed at midnight, not just the tidy code assigned to it, which is where the detail always hid. Three groups arrived at once, wanting different things.

Regulators moved first, and further than the industry noticed. Europe's medicines agency says AI may be used to handle side-effect reports and hunt for early warnings, provided a human stays in charge and the company is held accountable for the results. In June 2025, the FDA put its own AI, Elsa, on staff desks across the agency, and among its listed jobs is summarizing side effects for safety reviews. Read that twice. The regulator now reads at machine speed, and every company should assume its submissions are read that way. More quietly, regulators also built the comparison groups the idea depends on: American surveillance covering over a hundred million people, European, about a hundred and thirty million.

Then read the sequel. Within weeks, FDA staff told CNN that Elsa had confidently cited studies that did not exist. That is not an argument against the tool. It is the argument for the rule this article ends on.

Regulators aren't the only ones deploying AI in pharmacovigilance, see the specific tools already automating detection and reporting workflows.

→ Read: 7 AI Tools Revolutionizing Pharmacovigilance Workflows

Why Big Tech and Big Pharma Are Both In

Big tech's interest is simpler than it appears. NVIDIA built with Lilly the most powerful supercomputer any drug company owns, called by its healthcare chief "the new instrument of science." OpenAI went in with Sanofi. Anthropic released a version of Claude for life sciences. None of them wants to sell medicines. They want the corpus, because medicine is the largest body of unread writing left on earth.

Big pharma is paying because the arithmetic is brutal. A trial costs hundreds of millions and takes years, while safety data is already owned, already growing and already paid for. Novo Nordisk now drafts the regulatory reports that once took ten weeks in roughly ten minutes.

The One Rule That Holds It Up

One rule holds it up, and Elsa is why. Any machine entering a safety system must answer the question we put to a new colleague: " Show me where you got that. Point to the sentence. Name the source. A person still signs and answers for it. Our job is to make what they sign worth far more than it used to be.